Say AMD do provide significant value improvements with RDNA4 but don't see much uptick in sales/market share, do you think they stop that strategy and go back to mirroring nvidia's prices more closely (we tried and it didn't work so maximise $/card) or carry on for a few gens to see if they can get a foothold?
RDNA4
- Thread starter Frenetic Pony
- Start date
Bondrewd
Veteran
This post’s accuracy is not substantiated. Exercise discretion and wait for verified sources before accepting its claims.
they're not going back to anything, the strategy is to build a pile with hundreds of WGP and hopefully win it all!
(yes it's an exact copy of how CPU guys did it, but hey it works).
What was it 128megs for the 4k infinity cache ? I'm not sure that will be enough but we will have to see
For me to consider them i'd need 4080 performance at $600.
Bondrewd
Veteran
Yeah, but there are other caches besides MALL.
Slower but maybe (hopefully) for less.
they need to move units.
eh, might be too slow then compared to the competition. Certainly to slow for me. 4080 is the minimum i'd consider to replace my 3080 with. maybe it will last me another gen until rdna 5
that´s why I find hard to believe it will fit to ~240 mm2
Frenetic Pony
Veteran
I was thinking that too but looking into this, the fastest mass production GDDR6 is Samsung's 20gbps, SK Hynix and Micron are 18gbps. Samsung are sampling 24gbps and SK Hynix are (I assume) sampling 20gbps so it's not as bad as I first thought. Guess GDDR7 is around the corner so effort's put into that instead but would've been nice if 20gbps+ was more plausible for RDNA4
GDDR6 | DRAM | Samsung Semiconductor Global
Meet upgraded GDDR6's industry's fastest speed, greater power efficiency, and incredible versatility, with low voltage for high-bandwidth, high-computing tasks.semiconductor.samsung.com
GDDR6 part catalog
Explore Micron's GDDR6 part catalog, offering comprehensive information for high-performance memory solutions in graphics and gaming, ensuring reliability.www.micron.com
Plus side is more even arch/gen comparisons
The problem with GDDR7 for GPU designed with GDDR6 in mind would have a mismatch between bandwidth and compute, as well as memory support and compute size. Combined with a GPU that's going to be dropped in 18 months, and GDDR7 being more expensive per GB/s as it initially comes out, "eh forget it" sounds like a good strategy.
Bondrewd
Veteran
It's funny but not as eye popping as Zen5.
Mainstream parts will do just fine with G6.
https://videocardz.com/newz/amd-ryz...a3-5-cus-lp5x-8000-memory-and-32mb-mall-cache
it´s pitty that this cache didn´t make it´s way to the other Strix version. Still no mention about clock improvemnet (regarding to RDNA4)
AMD Ryzen 9050 “Strix Halo” specs leaked: 16 Zen5 cores and 40 RDNA3.5 CUs, LP5x-8000 memory and 32MB MALL cache
it´s pitty that this cache didn´t make it´s way to the other Strix version. Still no mention about clock improvemnet (regarding to RDNA4)
Last edited:
Frenetic Pony
Veteran
Is RDNA4 going to show up in laptops next year? Intel is starting to put up a fight for the iGPU space, and I'm sure AMD would like to continue its dominance of the iGPU gaming market. "Battlemage" is showing what I'm assuming to be a 40% uplift (it says 56 EU vs the 128Eu benchmarks, but I'm guessing they're going for the doubled up floating point pipeline thing just like everyone else). Admittedly just in Sisoft so not a great benchmark, but RDNA4 plus a 16mb LLC would certainly be welcome.
Bondrewd
Veteran
This post’s accuracy is not substantiated. Exercise discretion and wait for verified sources before accepting its claims.
No.
No they aren't.
That's very-very dead.
The next stop is Medusa, in 2026.
Frenetic Pony
Veteran
https://wccftech.com/amd-high-end-navi-4x-rdna-4-gpus-9-shader-engines-double-navi-31-rdna-3-gpu/
canceled Navi4X would have been a beast with 50% more shader engine than N31
canceled Navi4X would have been a beast with 50% more shader engine than N31
Last edited:
Bondrewd
Veteran
Navi4C and it was a weird new 9SE * 3SA * however much it had WGP setup.
PSA: Please be aware that any posts here are just unvalidated views when not accompanied with supporting evidence. There is no vetting process at B3D and anyone on the internet can post seemingly 'factual information', sometimes accruing an air of 'insider' through choice of language and writing style.
B3D is about the conversation. Engage in the conversation with a healthy degree of caution/scepticism when posts rely on a(n implied) position of authority rather than hard evidence. It is unhealthy to have ideas cemented around a given point of fact that isn't provable, and in the absence of proof, discussion should be open and consider all the (rational) possibilities.
B3D is about the conversation. Engage in the conversation with a healthy degree of caution/scepticism when posts rely on a(n implied) position of authority rather than hard evidence. It is unhealthy to have ideas cemented around a given point of fact that isn't provable, and in the absence of proof, discussion should be open and consider all the (rational) possibilities.
amd: add initial code for gfx12 (!29007) · Merge requests · Mesa / mesa · GitLab
Mesa 3D graphics library
gitlab.freedesktop.org
First set of GFX12 patches above for their open source drivers. At a quick glance there's super nothing interesting going on outside of how they emulate stream out/transform feedback with NGG on GFX12 compared to GFX11 ...
Assuming 64CUs at higher clocks is true then it would be quite the perf/CU regression to fall short of the 7800XT.
Personally, assuming that spec is true and clocks are a bit over 3Ghz I would expect an upper bound of 7900XT and a lower bound of 7900GRE with performance fluctuating between those bounds depending on if a game is more bandwidth or compute bound.
On the RT side I can't even begin to make any guesses, far too little for me to go on.
Bondrewd
Veteran
Well there are patches and kepler talked about them a bit.
Not here.
I have read those details, just don't have any kind of heuristic to guess what it does for RT performance besides the obvious intention that it will improve it from a PPA perspective if it goes as planned.
